AI Agent Model at 35B Rivals Trillion-Parameter AI

Agents-A1: a 35B open-source AI agent model

Shanghai AI Laboratory has released Agents-A1, a 35B agent model with a Mixture-of-Experts (MoE) design, as open source. In MoE, only part of the network's experts run for a given input. That makes inference cheaper than running a dense network of the same size.

Built by Shanghai AI Lab as a 35B MoE model

The paper's author line credits the Agents-A1 Team at Shanghai Artificial Intelligence Laboratory, which publishes under the org name InternScience. InternScience's own project page, separate from the paper, gives the same "35B MoE" description.

Released with open weights for broad use

The weights are released under the Apache 2.0 license and can be used broadly, including commercially. The code is also public on GitHub, so developers can run it and check the performance claim themselves.

How training on long trajectories scales the horizon

Parameter count is not the only driver of performance. How long a chain of steps a model was trained to handle also matters, and that is what this paper targets. Two models of the same size can still differ sharply in what they can carry through a long, multi-step task.

The agent horizon idea behind the paper's title

The title states the claim: rather than enlarging the model, it extends the agent's "horizon," so that a 35B MoE model aims for trillion-parameter-level performance. "Horizon" here means how long a sequence of steps the model can act over — how far ahead it can keep track of a task as it researches, reasons, and takes the next step, rather than answering in a single reply.

Long-horizon training instead of more parameters

Agents-A1 is trained on long knowledge-and-action trajectories that walk through an entire multi-step task, rather than short, one-off question-and-answer pairs. Learning whole workflows rather than isolated steps teaches the model to act with the next steps already in view. The repository README puts the average training trajectory at 45K tokens, and the goal is better performance on real tasks without a bigger model.

AI benchmarks against far larger trillion-parameter models

The paper compares Agents-A1 with 1T-parameter models such as Kimi-K2.6 and DeepSeek-V4-pro, and it leads on five benchmarks while staying close on three others. It does not match the larger models everywhere, so the split between "leading" and "competitive" matters.

Where the 35B model matches much larger AI

Agents-A1 leads on SEAL-0, IFBench, HiPhO, FrontierScience-Olympiad, and MolBench-Bind, and the abstract calls it highly competitive on SciCode, HLE, and BrowseComp. These numbers come from the authors, who credit horizon scaling rather than added size.

What the release means for builders

Agents-A1 offers an alternative to "make it bigger": train the model to look further ahead. If a smaller model can reach practical performance by learning long workflows, useful agents could run at much lower cost. For how smaller models hold up on your own hardware, see Local LLM Coding Now Hits 75% of Cloud AI.

Meituan took the opposite route on scale with a giant MoE model, covered in LongCat 2.0: Meituan's 1.6T Open MoE Coding Model. With Agents-A1's weights and code public under an open license, anyone can test the horizon idea on real tasks.